Energy Conversion and Management 52 (2011) 15–26
Contents lists available at ScienceDirect
Energy Conversion and Management
journal homepage: www.elsevier.com/locate/enconman
Probability distributions for offshore wind speeds
Eugene C. Morgan a,*, Matthew Lackner b, Richard M. Vogel a, Laurie G. Baise a
a
b
Dept. of Civil and Environmental Engineering, Tufts University, 200 College Ave., Medford, MA 02155, United States
Wind Energy Center, University of Massachusetts, Amherst, United States
a r t i c l e
i n f o
Article history:
Received 30 September 2009
Accepted 2 June 2010
Available online 14 July 2010
Keywords:
Wind speed
Probability distribution
Mixture distribution
Wind turbine energy output
Weibull distribution
Extreme wind
a b s t r a c t
In planning offshore wind farms, short-term wind speeds play a central role in estimating various engineering parameters, such as power output, extreme wind load, and fatigue load. Lacking wind speed time
series of sufficient length, the probability distribution of wind speed serves as the primary substitute for
data when estimating design parameters. It is common practice to model short-term wind speeds with
the Weibull distribution. Using 10-min wind speed time series at 178 ocean buoy stations ranging from
1 month to 20 years in duration, we show that the widely-accepted Weibull distribution provides a poor
fit to the distribution of wind speeds when compared with more complicated models. We compare distributions in terms of three different metrics: probability plot R2, estimates of average turbine power output, and estimates of extreme wind speed. While the Weibull model generally gives larger R2 than any
other 2-parameter distribution, the bimodal Weibull, Kappa, and Wakeby models all show R2 values significantly closer to 1 than the other distributions considered (including the Weibull), with the bimodal
Weibull giving the best fits. The Kappa and Wakeby distributions fit the upper tail (higher wind speeds)
of a sample better than the bimodal Weibull, but may drastically over-estimate the frequency of lower
wind speeds. Because the average turbine power is controlled by high wind speeds, the Kappa and Wakeby estimate average turbine power output very well, with the Kappa giving the least bias and mean
square error out of all the distributions. The 2-parameter Lognormal distribution performs best for estimating extreme wind speeds, but still gives estimates with significant error. The fact that different distributions excel under different applications motivates further research on model selection based upon
the engineering parameter of interest.
Ó 2010 Elsevier Ltd. All rights reserved.
1. Introduction
The offshore environment shows greater wind energy potential
than most terrestrial locations [1]. Accurate characterization of offshore winds will be essential in tapping this tremendous power
source. In wind turbine design and site planning, the probability
distribution of short-term wind speed becomes critically important in estimating energy production. Here short-term wind speed
U refers to mean wind speed averaged on the time scale of 10 min,
which is much longer than the time scale associated with turbulence and gusts. In engineering practice, the average wind turbine
power P w associated with the probability density function (pdf) of
wind speeds U is obtained from
b ¼
P
w
Z
1
Pw ðUÞf ðUÞdU;
ð1Þ
0
where f(U) is the pdf of U and where the turbines power curve Pw(U)
describes power output versus wind speed [2]. The largest uncer* Corresponding author. Tel.: +1 617 627 3098.
E-mail address: eugene.morgan@tufts.edu (E.C. Morgan).
0196-8904/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.enconman.2010.06.015
tainty in the estimation of P w lies in the choice of the wind speed
pdf f(U), since the turbine manufacturer knows Pw(U) fairly accurately. Minimizing uncertainty in wind resource estimates greatly
improves results in the site assessment phase of planning [3], and this
can be accomplished by utilizing a more accurate distribution (pdf).
By far the most widely-used distribution for characterization of
10-min average wind speeds is the 2-parameter Weibull distribution (W2) [2,4–9]. Sometimes the simple 1-parameter Rayleigh
(RAY) distribution offers a more concise fit to a sample, since it is
a special case of the W2 [2,10], but ultimately having only a single
model parameter makes the RAY much less flexible. Despite the
widespread acceptance of the W2 and RAY distributions, Carta
et al. [11] and others have noted that under different wind regimes
other distributions may fit wind samples better. Other distributions used to characterize wind speed include the 3-parameter
Generalized Gamma (GG), 2-parameter Gamma (G2), inverse
Gaussian, 2-parameter Lognormal (LN2), 3-parameter Beta, singly
truncated from below Normal, distributions derived from the Maximum Entropy Principle, and bimodal (two component mixture)
Weibull model (BIW). Kiss and Janosi [12] recommend the GG for
the ERA-40 dataset (6-hourly mean) after testing the RAY, W2
16
E.C. Morgan et al. / Energy Conversion and Management 52 (2011) 15–26
and LN2 distributions. Simiu et al. [13] use probability plot correlation coefficient (PPCC) goodness-of-fit tests to document that
the RAY distribution provides a poor approximation to hourly wind
speeds, but the authors do not advocate an alternative.
Recent studies show that Weibull mixture models (including
BIW) out-perform the W2 distribution [14–16]. However, these
studies only use a limited number of samples from confined geographic regions, and often explain the better fits of the mixture
models as unique exceptions to the rule. For example, Akpinar
and Akpinar [15] show that mixture models out-perform the Weibull distribution and other distributions derived from the Maximum Entropy Principle, but only use 4 samples from a single
region in Turkey. Carta and Ramirez [16] arrive at similar findings
for 16 samples in the Canary Islands, as do Jaramillo and Borja [14]
for one sample from La Ventosa, Mexico.
One goal of this study is to exploit a much larger and geographically diverse dataset than previous studies in order to evaluate
alternative wind speed distributional hypotheses, such as the
BIW model. We use data spread out over many different wind regimes, albeit all from ocean environments, in order to assess the
communicability of such models. While the offshore setting for
all samples in this study may limit the utility of some of our findings for onshore data, we find the use of this data essential in
removing topographic effects, which would complicate a similar
analysis performed on terrestrial wind speed samples.
Another primary goal of this study is to illustrate that finding an
optimal wind speed distributional model is contingent upon the
application of interest. Measures of goodness-of-fit commonly
used in wind literature (i.e. R2) [17,16,18,12,10,5,11] do not necessarily indicate how well a model predicts wind energy parameters.
Following Chang and Tu [19], a model validation approach based
on parameter estimation instead of R2 may find theoretical models
that are best suited for specific wind energy applications. In practice, one may resort to a variety of models to give optimal predictions of in situ conditions when lacking observations.
2. Data
The National Data Buoy Center (NDBC) [20] provides 10-min
mean wind speed observations recorded at 178 buoys around
North America (Fig. 1). We analyze the 10-min mean wind speed
data because the International Electrotechnical Commission requires the evaluation of this form of wind measurement in turbine
site assessment [21]. Depending on the type of buoy, the anemometers on the buoys measure wind speed at either 5 or 10 m above
sea level. Sample sizes (n) range from 4,314 to 889,699, corresponding to time spans of approximately 30 days to 20 years. This
dataset comprises the largest collection of first-order wind speed
observations that we could find readily available online and that
provides a relatively reliable and consistent measurement of
short-term wind speeds.
3. Wind speed distributions
We consider some of the most common wind speed distributions used in previous studies. We also consider numerous other
closely-related distributions, such as the Generalized Rayleigh
(GR), Kappa (KAP), and Wakeby (WAK) distributions, that have
not previously been investigated as wind speed models, but are
simple extensions or generalizations of commonly used models.
The 2-parameter GR distribution is a generalization of the 1parameter RAY distribution, and the KAP and WAK distributions
(with four and five parameters, respectively) are generalizations
of numerous other distributions.
Since the record lengths of the time series of wind speeds are
rather large, differences among parameter estimation methods will
not be nearly as important as differences among distributions. We
preferentially use maximum likelihood estimators (MLE) because
they usually yield lower mean square errors (MSE) associated with
model parameter estimates than method of moments (MOM) estimators for the large samples considered here. However, for some
distributions we use MOM, least squares (LS), or a parameter estimation method involving L-moments when MLEs are either too
cumbersome or unavailable, as discussed in each case. When using
MOM, we calculate the sample mean ðUÞ, standard deviation (S),
and skewness (G) as
n
1X
Ui ;
n i¼1
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
n
u 1 X
S¼t
ðU i UÞ2 ;
n 1 i¼1
P
1 ni¼1 ðU i UÞ3
G¼
;
n
S3
U¼
ð2Þ
ð3Þ
ð4Þ
where n is the number of observations in U, our random variable for
wind speed. The following describes each distribution in our study
and the parameter estimation methods used.
3.1. Rayleigh
The Rayleigh (RAY) distribution is the simplest distribution
commonly used to describe 10-min average wind speeds [2,4,17]
because it only has a single model parameter b. Its pdf f(U), and
cumulative distribution function (cdf) F(U), are
f ðU; bÞ ¼
U
2
b
exp !
1 U2
;
2 b2
ð5Þ
and
FðU; bÞ ¼ 1 exp !
1 U2
;
2 b2
ð6Þ
for U P 0 and b P 0. We estimate the single scale parameter b using
the MLE:
Fig. 1. Locations of buoys providing wind speed observations for this study. Data
source: http://www.ndbc.noaa.gov.
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
n
u1 X
^
b¼t
U2:
2n i¼1 i
ð7Þ
17
E.C. Morgan et al. / Energy Conversion and Management 52 (2011) 15–26
Often databases report wind speeds as two orthogonal components (u and v). If u and v are independent and identically distributed Gaussian random
variables with means of zero and standard
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
deviations of b2, U ¼ u2 þ v 2 follows a Rayleigh distribution [22].
However, NDBC does not report wind speeds in separate u, v
components.
3.4. 3-parameter Weibull
The 3-parameter Weibull (W3) is a generalization of the W2
distribution, where the location parameter s establishes a lower
bound (which was assumed to be zero for the W2 distribution).
Stewart and Essenwanger [25] found that the W3 fits wind speed
data better than the W2 model. The W3 pdf and cdf are
3.2. Weibull
The Weibull (W2) distribution constitutes the most widely accepted distribution for wind speed [2,4–9]. It is a generalization
of the RAY distribution, and has been shown to fit wind samples
better than RAY due to its more flexible form (with an additional
shape parameter) [17,12]. The Weibull pdf and cdf are
f ðU; b; aÞ ¼
bU b1
ab
" #
b
U
;
exp ð9Þ
i¼1
^
n
1 1X
ln U i ¼ 0
^ n
b
ð10Þ
i¼1
^ and then using
iteratively for the shape parameter estimate b,
n
1X
^
Ub
n i¼1 i
a^ ¼
ð18Þ
and
" b #
Us
;
FðU; b; a; sÞ ¼ 1 exp Pn
^
^b
i¼1 ðU i Þ lnðU i
Pn
^
^b
i¼1 ðU i Þ
s
^Þ
s
s
with U P 0. The RAY is a special case of the W2 with b = 2 [23].
MLEs are obtained by solving
U bi
a
ð19Þ
respectively, for U P s. We estimate the parameters using the modified MLEs recommended by Cohen [26]:
a
ab
ð8Þ
a
" #
b
U
FðU; b; aÞ ¼ 1 exp ;
n
P
" b #
Us
;
exp bU b1
a
and
Pn b^
i¼1 U i ln U i
f ðU; b; a; sÞ ¼
!1^
b
ð11Þ
n
1 1X
^Þ ¼ 0;
lnðU i s
^ n
b
ð20Þ
i¼1
!1^
n
b
1X
^
^Þb ;
ðU i s
n i¼1
!
a^
1
^
s þ 1 C 1 þ ^ ¼ U1 ;
b
nb^
a^ ¼
ð21Þ
ð22Þ
where n is the number of observations in sample U, and U1 indicates
^ is found by iteratively solving Eqs.
the minimum value of U. First, s
^ is found by iteratively solving Eq. (20). Finally,
(20)–(22), and then b
a^ is given by Eq. (21) using the estimates of the other two parameters [23,26].
^.
for the scale parameter estimate a
3.5. Lognormal
3.3. Generalized Rayleigh
The 2-parameter Lognormal (LN2) pdf and cdf are
Another generalization of the RAY distribution is the generalized Rayleigh (GR) model, which is sometimes referred to as the
2-parameter Burr Type X distribution [24]. While the GR has not
previously been applied to wind speed samples, it is a natural
extension of the RAY distribution (like the W2 distribution). It
has a pdf
f ðU; a; kÞ ¼ 2ak2 U exp½ðkUÞ2 ð1 exp½ðkUÞ2 Þa1 ;
ð12Þ
f ðU; l; rÞ ¼
"
#
1
ðlnðUÞ lÞ2
pffiffiffiffiffiffiffi exp
;
2r 2
rU 2p
ð23Þ
1 1
lnðUÞ l
pffiffiffi
;
þ erf
2 2
r 2
ð24Þ
and
FðU; l; rÞ ¼
and cdf
FðU; a; kÞ ¼ ð1 exp½ðkUÞ2 Þa ;
ð13Þ
for U > 0. We estimate the shape a and scale k parameters using the
modified moment estimators introduced by Kundu and Raqab [24]:
0
n
1X
U2 ;
n i¼1 i
ð14Þ
n
1X
U2 W 2 ;
V¼
n i¼1 i
ð15Þ
W¼
V
W
2
¼
^ þ 1Þ
w0 ð1Þ w0 ða
;
^ þ 1Þ wð1ÞÞ2
ðwða
ð16Þ
^ , and
which requires solving iteratively for a
^k ¼
where erf() is the error function from the Normal distribution, and
the parameters l and r are the mean and standard deviation of the
natural logarithm of U [5,12,17,27]. Since some observations may be
zero, we employ MOM estimators [28] instead of MLEs as follows:
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
^ þ 1Þ wð1Þ
wða
;
W
ð17Þ
0
where w() is the digamma function and w () is the polygamma function of order one.
1
U C
l^ ¼ ln B
@qffiffiffiffiffiffiffiffiffiffiffiffiffi2 A;
1 þ US 2
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
!
u
u
S2
t
r^ ¼ ln 1 þ 2 :
U
ð25Þ
ð26Þ
For both the 2- and 3-parameter Lognormal distributions, we
chose MOM over MLE because it showed considerably better fits
to the data and because it enables us to handle samples with zeros.
The results of Carta et al. [17] corroborate this statement with
short-term wind speed observations from the Canary Islands,
showing that MOM for LN2 fits the data better than MLE for 5 of
8 samples.
18
E.C. Morgan et al. / Energy Conversion and Management 52 (2011) 15–26
3.6. 3-parameter Lognormal
Hosking and Wallis [29] present this distribution, as well as an
explanation of the sample L-moments and their ratios.
The 3-parameter Lognormal (LN3) model is a shifted version of
the LN2, with pdf
"
#
2
1
ðlnðU fÞ lÞ
pffiffiffiffiffiffiffi exp
;
f ðU; l; r; fÞ ¼
2r 2
rU 2p
ð27Þ
and cdf
FðU; l; rÞ ¼
1 1
lnðU fÞ l
pffiffiffi
;
þ erf
2 2
r 2
ð28Þ
where the location parameter f establishes a lower bound on the
distribution. This distribution has yet to be fit to wind speeds, but
we expect it to perform as well or better than the LN2 model.
MOM estimates the three parameters via the following equations:
i13
1
1h
^ 2 ¼ 1 þ G2 þ Gð4 þ G2 Þ2
exp½r
2
i13
1
1h
þ 1 þ G2 Gð4 þ G2 Þ2
1;
2
!
1
S2
;
l^ ¼ ln
2
^
^ 2 Þ 1
2
exp ðr Þ½exp ðr
1 2
^f ¼ U exp l
^ ;
^þ r
2
ð29Þ
ð30Þ
f ðU; b; aÞ ¼ U
a1
exp Ub
;
ð41Þ
U
FðU; b; aÞ ¼ c a;
CðaÞ;
b
ð42Þ
ba CðaÞ
where c() is the lower incomplete gamma function [22]. In order to
keep null wind speeds in the sample, we estimate the shape a and
scale b parameters via MOM:
2
ð31Þ
^¼S ;
b
U
U2
a^ ¼ 2
S
ð43Þ
ð44Þ
as given by Carta et al. [17]. Again, we do not use the MLEs for this
distribution because they require removing all Ui = 0.
3.7. Generalized Normal
The Generalized Normal (GNO) model encompasses Lognormal
distributions with both positive and negative skewness, as well as
the Normal distribution [29]. We consider this pdf due to its more
general form than the LN2 pdf, which has previously been recommended for use with wind speed data [5,12,17,27]. The pdf and cdf
are
y¼
The 2-parameter Gamma (G2) distribution has been applied to
wind speed data by Sherlock [30], Kaminsky [27], and Auwera et al.
[31], for example. This distribution is often referred to as a Pearson
Type III, however in this paper we make a distinction between the
two distributions in order to evaluate the effect of the added location parameter which we include in the Pearson Type III (P3). The
G2 pdf and cdf are
and
which allow us to keep null values (U = 0) [28].
(
3.8. Gamma
3.9. Pearson type III
The Pearson type III (P3) distribution is a G2 distribution with a
third parameter for location, whose use for wind speed was initially advocated by Putnam [32]. Its pdf and cdf are
1
k lnð1 kðU nÞ=aÞ; k – 0
;
k¼0
ðU nÞ=a;
f ðU; a; k; nÞ ¼
expðky y2 =2Þ
pffiffiffiffiffiffiffi
;
a 2p
ð32Þ
ð33Þ
and
FðU; a; k; nÞ ¼ UðyÞ;
ð34Þ
where U() is the standard Normal cdf. The location n, scale a, and
shape k parameters relate to standard parameters for the LN2 and
LN3 distributions above by
k ¼ r;
ð35Þ
a ¼ r expðlÞ;
ð36Þ
and
n ¼ f þ expðlÞ;
ð37Þ
and we estimate these parameters using sample L-moments k1 and
k2 and the sample L-moment ratio s3:
2
4
6
^ s3 2:0466 3:6544s3 þ 1:8397s3 0:2036s3 ;
k
2
4
6
1 2:0182s3 þ 1:242s3 0:21742s3
^
^2 =2Þ
k k expðk
a^ ¼ 2
pffiffiffi ;
^
1 2U k= 2
^
^2 =2ÞÞ:
^n ¼ k1 a ð1 expðk
^
k
ð38Þ
ð39Þ
ð40Þ
f ðU; s; b; aÞ ¼ ðU sÞ
a1
exp Ub s
ba CðaÞ
;
ð45Þ
and
Us
FðU; s; b; aÞ ¼ c a;
CðaÞ:
b
ð46Þ
Again, to enable us to handle null wind speeds, we estimate the
location s, scale b, and shape a parameters using MOM [33]:
s^ ¼ U 2
S
;
G
^ ¼ SG ;
b
2
4
a^ ¼ 2 :
G
ð47Þ
ð48Þ
ð49Þ
The MLEs for this distribution require the removal of null wind
speeds [34].
3.10. Log Pearson type III
The log Pearson type III (LP3) distribution describes a random
variable whose logarithms follow a P3 distribution [33]. Thus, we
fit the LP3 distribution to the wind speed sample U by transforming U by the natural logarithm and fitting the P3 distribution. The
LP3 distribution is one of the most flexible 3-parameter distributions that exists. Thus, we sought to evaluate its performance, even
though we could not find any previous studies which evaluate its
use for wind speeds.
19
E.C. Morgan et al. / Energy Conversion and Management 52 (2011) 15–26
3.11. Generalized Gamma
The 3-parameter generalized Gamma (GG) distribution includes
the W2, P3, and other distributions as special cases, and has been
proposed as a wind speed model by several authors [12,17,31].
Its pdf and cdf are
f ðU; b; a; pÞ ¼ jpjU
pa1
p
exp Ub
pa
b CðaÞ
inally introduced by Harold Thomas [39,40] who lived on Wakeby
pond on Cape Cod. The pdf and cdf are not explicitly defined, and
the WAK distribution is most easily defined by its quantile
function:
a
F 1 ðFðUÞ; n; a; b; c; dÞ ¼ n þ ð1 ð1 FðUÞÞb Þ
b
c
ð1 ð1 FðUÞÞd Þ:
d
ð50Þ
;
and
p U
:
FðU; b; a; pÞ ¼ c a;
b
ð51Þ
In order to avoid removing null wind speeds from the samples
before fitting the model (as required by the MLEs), we estimate
the scale b and shape parameters a, p by MOM. First, we solve
^:
the following equation iteratively for a
^ Þ=ðw0 ða
^ ÞÞ3=2 ;
jGj ¼ w00 ða
ð52Þ
and then find the remaining parameters with the following
equations:
^ ÞÞ1=2 =S;
^ ¼ ðw0 ða
p
^ ¼ expðU wða
^ Þ=p
^Þ;
b
ð53Þ
Parameter estimation follows an algorithm outlined in Hosking
and Wallis [29] and Hosking [41] which relies on sample L-moments. We implement this distribution and its parameter estimation using the package lmomco [42] for the statistical software R
[43].
3.14. Bimodal Weibull mixture
The bimodal Weibull mixture distribution (BIW) is a mixture of
two W2 distributions. The BIW model exhibits two different modes
and has been recommended for use with wind speed data by Carta
et al. [17], Akpinar and Akpinar [15], Carta and Ramirez [18], and
Jaramillo and Borja [14]. The pdf and cdf are
f ðU; b1 ; a1 ; b2 ; a2 ; xÞ ¼ x
00
where w() is the digamma function, and w () and w () are the polygamma functions of order one and two [35].
The 4-parameter Kappa (KAP) distribution is a generalization of
many other distributions and includes as special cases the Generalized Logistic, Generalized Extreme Value, and Generalized Pareto
distributions [29]. Its pdf and cdf are
f ðU; n; a; k; hÞ ¼ a1 ð1 kðU nÞ=aÞ1=k1 ðFðUÞÞ1h ;
ð55Þ
and
FðU; n; a; k; hÞ ¼ ð1 hð1 kðU nÞ=aÞ1=k Þ1=h ;
ð56Þ
respectively. We estimate the location n, scale a, and two shape
parameters k, h by solving the following relationships between
these parameters and the sample L-moments k1, k2 and L-moment
ratios s3, s4:
^
^ ð1 g 1 Þ=k;
k1 ¼ ^n þ a
^
^ ðg g Þ=k;
k2 ¼ a
ð57Þ
ð58Þ
2
s3 ¼ ðg 1 þ 3g 2 2g 3 Þ=ðg 1 g 2 Þ;
s4 ¼ ðg 1 þ 6g 2 10g 3 þ 5g 4 Þ=ðg 1 g 2 Þ;
ð59Þ
ð60Þ
where
gr ¼
8
^ Cðr=hÞ
^
r Cð1þkÞ
>
< h^1þk^ Cð1þkþr=
^
^ ;
hÞ
^>0
h
^
^
^
>
^<0
: rCð1þ1þkÞ^kCðkr=hÞ ; h
^
^
ðhÞ
ab11
þ ð1 xÞ
" #
b
U 1
exp a1
b2 U
b2 1
ab22
" #
b
U 2
;
exp a2
ð63Þ
and
3.12. Kappa
1
b1 U b1 1
ð54Þ
0
ð62Þ
:
ð61Þ
Cð1r=hÞ
^ and
Given s3 and s4, we solve Eqs. (57) and (58) iteratively for k
^ [29].
h
3.13. Wakeby
The 5-parameter Wakeby (WAK) distribution is a generalization
of many existing pdfs (including LN2 and P3), but the WAK pdf can
exhibit shapes that other distributions cannot mimic [36]. It has
been considered as a super population or parent distribution in
many previous studies in the earth sciences [37,38] and was orig-
" #!
b
U 1
FðU; b1 ; a1 ; b2 ; a2 ; xÞ ¼ x 1 exp a1
" #!
b
U 2
þ ð1 xÞ 1 exp ;
a2
ð64Þ
where the shape b and scale a parameters have subscripts corresponding to the two different modes, and the mixing parameter x
describes the proportion of observations belonging to the first component. We first estimate b1 and a1 using the MLE for the W2, and
assign b2 = 1, a2 = 1, and x = 0.5. Then, using these estimations as
our initial parameters, we optimize the five parameters to give
the LS fit of the model cdf to the empirical (wind sample) cdf. The
least squares method is usually applied to the cdf [17].
4. Results and discussion
We evaluate the goodness-of-fit of the various fitted distributions to the buoy wind speed samples using the coefficients of
determination associated with the resulting P–P probability plots,
an example of which is presented in Fig. 2. R2 is used widely for
goodness-of-fit comparisons and hypothesis testing because it
quantifies the correlation between the observed probabilities and
the predicted probabilities from a distribution. Here we employ
P–P probability plots instead of Q–Q probability plots (which plot
predicted quantiles versus observed quantiles) because P–P plots
do not require specially designed quantile unbiased plotting positions for each pdf considered. Instead, we use the Weibull plotting
position F(U) = i/(n + 1), where i=1,. . ., n, in all cases because it always gives an unbiased estimate of the observed cumulative probabilities regardless of the underlying distribution considered, and
does not estimate the highest observed wind speed as the maximum possible wind speed (i.e. F(Un) – 1).
A larger value of R2 indicates a better fit of the model cumulative probabilities b
F to the empirical cumulative probabilities F.
We determine R2 as:
E.C. Morgan et al. / Energy Conversion and Management 52 (2011) 15–26
(c) of the W2 distribution are obtained solely in terms of the shape
parameter b from the following relations [23]:
1.0
20
R^2 = 0.9882
0.6
Cv ¼
0.4
c¼ 0.0
0.2
0.4
0.6
0.8
2
;
ð66Þ
Pn
2
b
i¼1 ð F i FÞ
Pn
2
FÞ þ i¼1 ðF i
b
F i Þ2
ð65Þ
;
P
where F ¼ 1n ni¼1 b
F are
F i . The estimated cumulative probabilities b
obtained from the various models’ cdfs described previously. We
chose R2 as the goodness-of-fit measure for ease of comparison between distributions, and also because of its widespread use in similar studies [5,17,44].
Fig. 3 ranks the distributions (from left to right) in terms of their
goodness-of-fit as measured by their median R2 values. Although
the W2, the most widely accepted distribution for modeling wind
speeds, performs the best out of all 2-parameter distributions and
even better than some three-parameter distributions, it is not the
model with the best goodness-of-fit in terms of R2. The BIW, KAP,
and WAK all show significant improvement in fit over the other pdfs,
with the BIW exhibiting the highest R2 values and the highest median R2. In comparison to the WAK, the BIW model exhibits a wider
variability in R2 values and lower R2 values for the poorly fit sites.
4.1. The probability distribution of offshore wind speeds
The poor performance of the W2 distribution for these samples
compared to some of the more complex models can be further explained using moment diagrams as shown in Fig. 4. The theoretical
relationships for coefficient of variation ðC v ¼ S=UÞ and skewness
3
2 3=2
2 3C 1 þ 1b C 1 þ 2b C 1 þ 1b
:
2 3=2
C 1 þ 2b C 1 þ 1b
1.0
Fig. 2. Example P–P probability plot from buoy 41040. Red line denotes a 1:1
relationship between the model-predicted probabilities and the observed probabilities (i.e. R2 = 1). (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)
b C 1 þ 2b C 1 þ 1b
Observed Probability
i¼1 ð F i
C 1 þ 1b
C 1 þ 3b C 1 þ 1b
0.0
R2 ¼ Pn
C 1 þ 2b C 1 þ 1b
and
0.2
W2 Probability
0.8
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð67Þ
Here moment diagrams are constructed for both synthetic and
empirical samples. Fig. 4 illustrates that the coefficient of variation
^j
^ j , and b
and skewness for synthetic W2 samples with the same nj, a
as the buoy samples closely mimic the W2 theoretical relationship,
whereas the moments of the actual buoy samples do not. Fig. 4
provides a clear documentation of how poorly the tail behavior
of the empirical samples is mimicked by the W2 distribution.
The scatter about the theoretical relationships in Fig. 4b is shown
to arise from real distributional departures alone, and not sampling
variability.
While we cannot compare the buoy sample moments and BIW
synthetic sample moments to a theoretical BIW moment relationship (because theoretical moment relationships are incalculable for
this model), we can compare BIW R2 values for the buoy samples
(from Fig. 3) to BIW R2 values for synthetic BIW samples. Fig. 5 provides a comparison of the goodness-of-fit of the W2 and BIW distributions to both the real buoy samples and the synthetic buoy
samples. For each real buoy sample, we randomly generate 2 synthetic samples: one from the W2 model and one from the BIW
model. We use the same parameters and sample sizes as the actual
samples, and then fit the two distributions to their respective synthetic samples using the same parameter estimation methods described previously. The W2 and BIW models fit the synthetic W2
and BIW samples almost exactly, as expected, with any departures
of R2 values away from 1 reflecting sampling error alone. If the
buoy samples truly came from a W2 or BIW parent distribution,
they would have R2 values comparable to those of the W2 or
BIW synthetic samples.
1.00
0.99
R2
0.98
0.97
0.96
0.95
G2
2
LN2 RAY
LP3
GR
LN3
GG
PE3
W2
GNO
W3
WAK KAP BIW
Fig. 3. Boxplots of R values for the distribution fits to the buoy wind speed samples. Distributions are ordered from left to right by ascending median R2 value (heavy black
line in each box).
21
E.C. Morgan et al. / Energy Conversion and Management 52 (2011) 15–26
Real Buoy Samples
0.7
0.7
0.6
0.6
Cv
Cv
Synthetic Weibull Samples
0.5
0.4
0.5
0.4
Syn. Weibull
Theoretical Weibull
0.3
−0.5
0.0
0.5
1.0
Buoy Wind Speeds
Theoretical Weibull
0.3
1.5
−0.5
0.0
0.5
Skewness
1.0
1.5
Skewness
Fig. 4. Comparison of moments between synthetic W2 samples (left) with the real buoy samples (right). The red line represents the theoretical relationship between
moments for the W2 distribution. Note the wide dispersion of points around the red line for the buoy samples versus the proximity to the red line that the synthetic samples
have. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
1.000
0.995
0.995
0.990
0.990
0.985
0.985
0.980
0.980
R2
1.000
a
0.975
b
0.975
W2 Real
W2 Synthetic
BIW Real
BIW Synthetic
Fig. 5. Boxplots of R2 values for (a) W2 fits to the real buoy samples (left) and to synthetic samples of same size and parameters (right), and (b) BIW fits to the real buoy
samples (left) and to synthetic samples of same size and parameters (right). The greater similarity between the real and synthetic BIW R2 values (versus those for W2)
illustrates that the buoy wind speed samples are better modeled with a BIW distribution.
4.2. BIW mixing parameter
Further comparison of the W2 and BIW fits can be found in the
mixing parameter x of the BIW distribution, which tells us about
the significance of the second mode in the buoy samples. x = 0.5
indicates that the two Weibull modes are comparable in size,
whereas a value of 1 indicates that no second mode exists. We constrain x between 0.5 and 1 because we interpret it as the weight of
the ‘‘strongest” mode in the model. That is, a BIW model with
x < 0.5 can be considered as an alternative model with xalt = 1 x
and with the shape and scale parameters switched between the
two modes (i.e. b1, a1, b2, and a2 for the initial model respectively
become b2,alt, a2,alt, b1,alt, and a1,alt for the alternative model).
Fig. 6 illustrates the distribution of mixing parameters for the
BIW fits to the buoy samples. 68 of the 178 samples have x = 1,
which implies that they follow a W2 distribution. For these 68
100
68 values (out of 178) = 1
80
Frequency
The departures of the R2 values from unity for the W2 real samples are much too great to be explained by random sampling error
alone (Fig. 5a). The discrepancy between the two boxplots in Fig. 5a
results from the inadequacy of the W2 distribution to accurately
model the buoy wind speed samples, and cannot be explained by
random sampling error. We observe that the BIW model does a
much better job at fitting the buoy wind speed samples (Fig. 5b)
than the W2 model.
60
40
20
0
0.5
0.6
0.7
0.8
0.9
1.0
BIW mixing parameter value
Fig. 6. Distribution of BIW mixing parameters x from fits to buoy samples. Values
of 1 indicate W2 distributed samples.
samples, any improvement in R2 from W2 to BIW (e.g. Fig. 7, rightside histogram) only highlights the benefit of using least squares
parameter estimation instead of MLE for W2. The remaining samples
22
E.C. Morgan et al. / Energy Conversion and Management 52 (2011) 15–26
display a wide range of bimodality, with x having approximately
uniform frequency from 0.5 to 0.95 (Fig. 6). Forty-eight samples have
close to 1 (that is 0.95 6 x < 1), indicating diminutive second modes.
Thus, a little more than 1/3 of the samples (62 samples) show varying
degrees of bimodality (0.5 6 x < 0.95).
Mapping x indicates some general spatial structure in the degree of bimodality of the wind speed distributions (Fig. 7; see
http://engineering.tufts.edu/cee/geohazards/peopleMorgan.asp for
a Google Earth file showing plots for all buoys). We can see that
the California coast has a majority of low x values, where the
BIW substantially improves the fit over W2, but higher values of
x also exist in that region. The east coast of North America shows
130˚W
125˚W
120˚W
predominantly high values of x (with little difference between the
BIW and W2 fits), with lower values interspersed. Thus, no truly
homogenous region exists, and one cannot say for which regions
to use which distribution. Therefore, if one is concerned with overall goodness-of-fit, the conservative practice would be to use the
BIW as the default wind speed distribution.
4.3. Model performance in terms of average power output
Here we evaluate how well each fitted pdf performs in terms of
its ability to agree with the empirical average wind turbine power
output computed using
85˚W
115˚W
80˚W
75˚W
70˚W
45˚N
Buoy 44028
0.12
W2 , R^2 = 0.9985
BIW , R^2 = 0.9993
Buoy 46063
0.04
0.00
Density
0.08
0.12
W2 , R^2 = 0.9862
BIW , R^2 = 1.0000
35˚N
0
5
10
15
Wind Speed (m/s)
20
0.00
0.04
40˚N
40˚N
Density
0.08
45˚N
0
35˚N
5
10
Wind Speed (m/s)
30˚N
15
Mixing Parameter
0.5 ≤ ω < 0.7
0.7 ≤ ω < 0.9
25˚N
0.9 ≤ ω ≤ 1.0
130˚W
125˚W
120˚W
85˚W
80˚W
75˚W
Fig. 7. Map of BIW mixing parameters. Histograms with fitted W2 and BIW distributions exemplify two mixing parameter end-member cases. Left: California sample with a
much better fit given by the BIW model (x = 0.53). Right: Massachusetts buoy data with little difference between W2 and BIW (x = 1).
% Error in Pw
10
5
(a)
0
−5
−10
GNO
% Error in Pw
10
5
PE3
G2
LN2
RAY
W3
W2
LP3
GG
GR
BIW
LN3
WAK
KAP
LN2
PE3
G2
BIW
LP3
W3
LN3
RAY
W2
GR
GG
WAK
KAP
(b)
0
−5
−10
GNO
Fig. 8. Boxplots of percent error between time series (empirical) and model-predicted average wind machine power output for each distribution. (a) Ranks by descending
mean square error from left to right, and (b) ranks by descending bias.
23
E.C. Morgan et al. / Energy Conversion and Management 52 (2011) 15–26
100
0
15
Wind Speed (m/s)
20
600
300
100
0
300
200
Power (KW)
0.08
100
0
5
10
15
20
Buoy 46081
600
400
Power (KW)
W2 , P = 130
BIW , P = 141.1
KAP , P = 142.3
WAK , P = 139.6
P(U) , P = 143.1
0.00 0.05 0.10 0.15 0.20 0.25 0.30
200
300
400
500
600
0.20
0.15
0.10
10
Buoy 46014
f
0.00
5
30
0.06
0
Buoy cdrf1
0.05
Density
0
25
W2 , P = 156.3
BIW , P = 162.8
KAP , P = 164.9
WAK , P = 165
P(U) , P = 165.1
20
W2 , P = 48.2
BIW , P = 41.2
KAP , P = 47.7
WAK , P = 47.8
P(U) , P = 47.5
20
200
15
e
15
0
c
10
10
0.10
0.12
300
200
100
0
0.00
5
5
0.04
600
500
400
0.15
0.10
0.05
Density
0
Buoy 51001
W2 , P = 196.6
BIW , P = 198.7
KAP , P = 203.8
WAK , P = 204.4
P(U) , P = 203.7
0
0.00
15
0.02
b
10
0.00
5
Power (KW)
0.10
200
200
100
0
0.00
0
500
0.15
600
500
400
W2 , P = 187.4
BIW , P = 181.9
KAP , P = 187.1
WAK , P = 187.6
P(U) , P = 185.9
300
0.10
0.15
W2 , P = 59.6
BIW , P = 53.3
KAP , P = 59
WAK , P = 58.8
P(U) , P = 58.7
0.05
Density
Buoy 42057
d
400
Buoy 46086
a
600
for each buoy sample with size n. The model estimates of the power
b are obtained from Eq. (1), and we use the regression
output P
w
equation for the Vestas V47-666 KW turbine power curve Pw(U)
provided by Chang and Tu [19]. Because turbines do not exist at
any of our buoy locations, and because we generally have large
samples in this study, we assume the empirical time series estimate
as our ‘‘true” capacity factor. Chang and Tu [19] find that the time
series approach generally estimates capacity factors closer to the
actual measured (from the turbines) capacity factors than the distributional approach.
500
ð68Þ
400
n
1X
Pw ðU i Þ
n i¼1
0.05
Pw ¼
0
5
10
15
20
Wind Speed (m/s)
Fig. 9. Histograms with fitted W2, BIW, and WAK pdfs. Power curves P(U) are inserted only to illustrate that higher wind speeds matter most for estimating average power
output.
24
E.C. Morgan et al. / Energy Conversion and Management 52 (2011) 15–26
Fig. 8 presents the percent error associated with the average
power estimates for all the pdfs considered previously. Here the
percent error is defined as
b P P
w
w
100%:
Pw ð69Þ
Because the power curve weights the influence that probabilib , we see a different performance
ties of higher speeds have on P
w
of the distributions in Fig. 8 than we did in Fig. 3, which ranks
based on an unweighted goodness-of-fit to all wind speeds in a
sample. In Fig. 8a, we rank the distributions by
P
b P Þ2 , where N = 178 is the number of buoy
MSE ¼ N1 Nj¼1 ð P
w;j
w;j
P
b P Þ.
samples, and in Fig. 8b we rank by bias ¼ N1 Nj¼1 ð P
w;j
w;j
Fig. 8 illustrates that the BIW model produces average power estimates with lower MSE than the W2 model (16.8 versus 34.8), but
these estimates have more bias (3.6 versus 1.8). It should be
noted here that BIW under-predicts the mean power at every buoy.
Also note here that the W2 model does not perform the best out of
the 2-parameter models, like it did when comparing R2. Instead,
the GR distribution gives slightly better estimates of average power
output than the W2 distribution.
Fig. 8 shows that the KAP distribution yields the least MSE and
bias in estimation of mean power output (0.3 and 0.1, respectively), followed closely by the WAK model (0.4 and 0.2). Inspecting the fits of these two distributions to the wind speed histograms
in Fig. 9, we begin to understand why the KAP and WAK perform
worse than BIW in terms of R2 and better than BIW in terms of
power output. While the BIW provides a good fit to the entire histogram, the KAP and WAK models fit better at the upper tail (i.e.,
b : 4–25 m/s for this power
the wind speeds that matter most in P
w
curve), but then diverge from the samples at low wind speeds.
The WAK distribution grossly over-predicts the frequency of null
wind speeds in every case and sometimes over-predicts the highest density (e.g., Fig. 9a–d). The KAP either over-predicts (Fig. 9a,
c and d) or under-predicts (Fig. 9b and e) the peak probability density, and sometimes gives wildly inaccurate null value densities
(Fig. 9f), which accounts for some of the lower KAP R2 values. Thus,
b
the agreement between the models’ average power estimates P
w
and the empirical average power estimates Pw depends on how
well the models fit the wind speeds between the cut-in and cutout values (4 m/s and 25 m/s, respectively) of the power curve,
and this predictive capability does not necessarily correspond to
the models’ R2 values.
4.4. Model performance in terms of extreme wind speeds
In order to evaluate how well the models estimate extreme
wind speeds, we compare the number of observations that each
model predicts will exceed various percentiles (total exceedances
x) to the expected number of exceedances E[x], an analysis based
on simple probability theory and previously applied to flood flows
[45]. For various low probabilities p = 1.9 105, 1.9 106,
1.9 107, we generate model-predicted extreme events from
the quantile functions F1(1 p), where the p is the probability
that the event will be exceeded by any observation Ui. We chose
these particularly low p values because with such large sample
sizes, we want to examine sufficiently rare events. Predicting extreme wind speeds plays an important role in turbine design,
1000000
500000
p = 1.9e−05
E[x] = 1375
100000
50000
10000
5000
a
1000
|E[x] − Total Exceedances (x)|
LP3
GG
KAP
WAK
BIW
RAY
W2
W3
GR
GNO
PE3
LN3
LN2
1000000
p = 1.9e−06
E[x] = 137
100000
10000
1000
b
100
LP3
GG
KAP
WAK
BIW
RAY
W2
W3
GR
GNO
PE3
LN3
LN2
p = 1.9e−07
E[x] = 13
100000
1000
c
10
LP3
GG
KAP
WAK
BIW
W2
RAY
W3
GR
GNO
PE3
LN3
LN2
Fig. 10. Total number of observations from all 178 buoys exceeding the (100 p)th percentile of each fitted distribution. a, b, and c correspond to p = 1.9 105, 1.9 106,
1.9 107, respectively. Distributions are ranked from left to right by increasing agreement with E[x] (i.e. decreasing jE[x] Total exceedances (x)j).
E.C. Morgan et al. / Energy Conversion and Management 52 (2011) 15–26
where one normally considers the 50-year return gust value. Traditionally determined via annual maximum data, this extreme gust
value can be computed as 1.4 times the 50-year return event determined from the distribution of 10-min mean values [4].
For choosing a suitable model for estimation of extreme design
events, such as the wind speed with an average return period of 50
years, we recommend fitting extreme value distributions, such as
the Generalized Extreme Value distribution to series of annual
maximum wind speeds. We have not considered the prediction
of such events here. Instead, in this section, we simply evaluate
the ability of various models to capture the extreme tails of 10min wind speeds. Such analyses can be related to annual frequency
analyses and design return period estimation, but we have not addressed those issues here.
We count the number of observations Ui in each sample exceeding the extreme quantiles generated by each model. These quantiles are generated for each sample and each distribution using
the fitted parameters from the previous sections. For each distribution, we sum these exceedances over all samples to give the total
number of exceedances x used in Fig. 10. The expected number
of exceedances is E[x] = mp for each p, where m is the total number
of 10-min observations (72 million, or 1376 site-years). Thus,
we expect higher wind speed events (lower p) to be exceeded fewer times and correspond to a decrease in E[x].
Fig. 10 compares the absolute difference between the distributions total number of exceedances x and E[x] because the best
model for estimating an extreme wind speed from 10-min average
data should be in best agreement with E[x]. We see that for all p, all
distributions perform roughly the same relative to each other, with
the LN2 distribution performing the best. However, the errors
(jE[x] xj) associated with the LN2 model are still significant
(1148, 78, and 15 exceedances, for Fig. 10a, b, and c, respectively;
which come out to 83%, 57%, and 115% of their respective E[x] values). Unfortunately, constructing confidence intervals for the estimated number of exceedances as Vogel et al. [45] have done is
challenging in our case because our observations cannot be considered independent.
5. Conclusion
In terms of characterizing entire wind speed samples, the BIW,
KAP, and WAK distributions offer significantly better fits than other
models, with the BIW giving the highest R2 values. These models
are all quite complex, with each having 4–5 parameters. Considering only the simpler models (2–3 parameters), the traditional W2
distribution generally gives larger R2 values than all the 2-parameter models, and some of the 3-parameter models. Using many
large samples spread over a wide geographic region, we show that
the superior fit of the BIW to the data (versus the W2 distribution)
cannot be explained by sampling error. BIW synthetic samples
more closely resemble the real data than W2 synthetic samples.
The BIW model led to better goodness-of-fit than the W2 at all
locations, to varying degrees. Mapping the BIW mixing parameter
shows the degree to which the BIW differs from the W2 at all locations, and we could not discern any regions showing homogeneity
of the mixing parameter.
When predicting power output, the BIW no longer performs as
well as other distributions. Because the BIW model does not necessarily model the wind speeds between the cut-in and cut-out values of the power curve as well as the KAP or WAK, it led to
considerably greater errors in average wind turbine power output
estimates. Among all the models considered, the KAP distribution
predicted average power output with the lowest MSE and bias.
Interestingly, the LN2 distribution yielded the best estimate of extreme wind speeds, but still exhibited large errors.
25
The fact that different distributions perform better for different
metrics suggests that one’s selection of model depends on the wind
energy application. While the KAP model best estimates average
power output, there may be yet another distribution that is best
when estimating fatigue loads, for example. If one’s goal is to obtain that distribution which performs best under a particular application, further analyses are needed of the type performed here that
validate models based on average power output and extreme wind
speed.
Acknowledgements
We thank the Wittich Foundation for supporting this research
through the Peter and Denise Wittich Family Fund for Alternative
Energy Research at Tufts University. Additional funding was provided by NSF Grant OISE-0530151.
References
[1] US Dept. of Energy, Wind powering America; September 2009. <http://
www.windpoweringamerica.gov/>.
[2] Manwell JF, McGowan JG, Rogers AL. Wind energy explained: theory, design
and application; 2002.
[3] Lackner MA, Rogers AL, Manwell JF. Uncertainty analysis in MCP-based wind
resource assessment and energy production estimation. In: Journal of solar
energy engineering – transactions of the ASME: AIAA 45th aerospace sciences
meeting and exhibit, vol. 130; 2008.
[4] Burton T, Sharpe D, Jenkins N, Bossanyi E. Wind energy handbook. Wiley; 2001.
[5] Garcia A, Torres JL, Prieto E, Francisco AD. Fitting wind speed distributions: a
case study. Solar Energy 1998;62(2):139–44.
[6] Harris RI. Generalised pareto methods for wind extremes. useful tool or
mathematical mirage? J Wind Eng Ind Aerodyn 2005;93(5):341–60.
doi:10.1016/j.jweia.2005.02.004.
[7] Harris RI. Errors in gev analysis of wind epoch maxima from Weibull parents.
Wind Struct 2006;9(3):179–91.
[8] Ramirez P, Carta JA. Influence of the data sampling interval in the estimation of
the parameters of the Weibull wind speed probability density distribution: a
case study. Energy Convers Manage 2005;46(15–16):2419–38.
[9] Hennessey JP. Some aspects of wind power statistics. J Appl Meteorol
1977;16(2):119–28.
[10] Celik AN. A statistical analysis of wind power density based on the Weibull and
Rayleigh models at the southern region of Turkey. Renew Energy
2004;29(4):593–604.
[11] Carta JA, Ramirez P, Velazquez S. Influence of the level of fit of a density
probability function to wind-speed data on the WECS mean power output
estimation. Energy Convers Manage 2008;49(10):2647–55.
[12] Kiss P, Janosi IM. Comprehensive empirical analysis of ERA-40 surface wind
speed distribution over Europe. Energy Convers Manage 2008;49(8):2142–51.
doi:10.1016/j.enconman.2008.02.003.
[13] Simiu E, Heckert N, Filliben J, Johnson S. Extreme wind load estimates based on
the Gumbel distribution of dynamic pressures: an assessment. Struct Safety
2001;23(3):221–9.
[14] Jaramillo OA, Borja MA. Wind speed analysis in La Ventosa, Mexico: a bimodal
probability distribution case. Renew Energy 2004;29(10):1613–30.
[15] Akpinar S, Akpinar EK. Estimation of wind energy potential using finite
mixture distribution models. Energy Convers Manage 2009;50(4):877–84.
[16] Carta JA, Ramirez P. Use of finite mixture distribution models in the analysis of
wind energy in the Canarian Archipelago. Energy Convers Manage
2007;48(1):281–91.
[17] Carta JA, Ramirez P, Velazquez S. A review of wind speed probability
distributions used in wind energy analysis Case studies in the Canary
Islands. Renew Sustain Energy Rev 2009;13(5):933–55.
[18] Carta JA, Ramirez P. Analysis of two-component mixture Weibull statistics for
estimation of wind speed distributions. Renew Energy 2007;32(3):518–31.
[19] Chang T-J, Tu Y-L. Evaluation of monthly capacity factor of WECS using
chronological and probabilistic wind speed data: a case study of Taiwan.
Renew Energy 2007;32(12):1999–2010.
[20] NDBC. Noaa’s national data buoy center; 2009. <http://www.ndbc.noaa.gov/>
[07.01.09].
[21] IEC. Wind turbines – part 1: design requirements. Tech. rep. 61400-1 Ed.3,
International Electrotechnical Commission; 2005.
[22] Krishnamoorthy
K.
Handbook
of
statistical
distributions
with
applications. New York: Chapman & Hall; 2006.
[23] Murthy DNP, Xie M, Jiang R. Weibull models 2004.
[24] Kundu D, Raqab M. Generalized Rayleigh distribution: different methods of
estimations. Comput Stat Data Anal 2005;49(1):187–200. doi:10.1016/
j.csda.2004.05.008.
[25] Stewart DA, Essenwanger OM. Frequency-distribution of wind speed nearsurface. J Appl Meteorol 1978;17(11):1633–42.
26
E.C. Morgan et al. / Energy Conversion and Management 52 (2011) 15–26
[26] Cohen AC. Multi-censored sampling in 3 parameter Weibull distribution.
Technometrics 1975;17(3):347–51.
[27] Kaminsky FC. Four probability densities (log-normal, gamma, Weibull, and
Rayleigh) and their application to modelling average hourly wind speed. In:
Proceedings of the international solar energy society; 1977. p. 19.6–19.10.
[28] Stedinger JR. Fitting log normal-distributions to hydrologic data. Water Resour
Res 1980;16(3):481–90.
[29] Hosking J, Wallis J. Regional frequency analysis: an approach based on Lmoments. Cambridge University Press; 1997.
[30] Sherlock RH. Analyzing winds for frequency and duration. Meteorol Monogr
1951;1:42–9.
[31] Auwera L, Meyer F, Malet L. The use of the Weibull three-parameter model for
estimating mean wind power densities. J Appl Meteorol 1980;19:819–25.
[32] Putnam PC. Power from the wind. New York: D. Van Nostrand Company, Inc.;
1948.
[33] Griffis VW, Stedinger JR. Log-Pearson type 3 distribution and its application in
flood frequency analysis. II: parameter estimation methods. J Hydrol Eng
2007;12(5):492–500.
[34] Hirose H. Maximum likelihood parameter estimation in the three-parameter
gamma distribution. Comput Stat Data Anal 1995;20(4):343–54.
[35] Stacy EW, Mihram GA. Parameter estimation for a generalized Gamma
distribution. Technometrics 1965;7(3):349–58.
[36] Houghton JC. Birth of a parent – Wakeby distribution for modeling flood flows.
Water Resour Res 1978;14(6):1105–9.
[37] Vogel RM, Mcmahon TA, Chiew FHS. Floodflow frequency model selection in
Australia. J Hydrol 1993;146(1–4):421–49.
[38] Guttman NB, Hosking JRM, Wallis JR. Regional precipitation quantile values for
the continental United States computed from L-moments. J Climate
1993;6(12):2326–40.
[39] Landwehr JM, Matalas NC, Wallis JR. Estimation of parameters and quantiles of
Wakeby distributions. 1. Known lower bounds. Water Resour Res
1979;15(6):1361–72.
[40] Landwehr JM, Matalas NC, Wallis JR. Estimation of parameters and quantiles of
Wakeby distributions. 2. Unknown lower bounds. Water Resour Res
1979;15(6):1373–9.
[41] Hosking JRM. Fortran routines for use with the method of L-moments, version
3. Research report RC 20525, IBM Research Division; 1996.
[42] Asquith WH. Lmomco: L-moments, trimmed L-moments, L-comoments, and
many distributions, R package version 0.96.3; 2009.
[43] R Development Core Team. R: a language and environment for statistical
computing, R Foundation for Statistical Computing, Vienna, Austria; 2009.
<http://www.R-project.org>.
[44] Ramirez P, Carta JA. The use of wind probability distributions derived from the
maximum entropy principle in the analysis of wind energy: a case study.
Energy Convers Manage 2006;47(15–16):2564–77.
[45] Vogel RM, Thomas WO, Mcmahon TA. Flood-flow frequency model selection in
Southwestern United States. J Water Resour Plan Manage – ASCE
1993;119(3):353–66.